CRAFT_ANATOMY_STANDARDS.md

pipelines/CRAFT_ANATOMY_STANDARDS.md

CRAFT ANATOMY STANDARDS — the character-artist bar, and the calibration tables the anatomy bar reads

PROPOSAL TIER. Created 2026-08-08. This file did not exist before; nothing is superseded by it. It is

the craft half of a pair. The anthropometric half already exists and is good —

build/3d/characters/body_specs/AGE_PROPORTION_RESEARCH.md — and this document does not repeat it,

does not re-derive it, and does not contradict it. Where the two touch, that document wins on

measured anthropometry and this one adds the artist-side layer it does not carry: what a character

artist actually looks at, in what order, against which named landmarks, and what separates a mesh

that is *numerically* on target from one that reads as a person.

Occasion: a generated A14 mesh shipped gaunt and deformed through every armed gate, and Josh's read

was one line — *"not anatomically correct."* Numbers alone did not catch it. §7 explains, with our own

arithmetic, exactly why they could not, and the tables in §3–§5 are written to be lifted directly into

harness/asset_factory/anatomy_bar.py.

Citation audit, 2026-08-08 — read this before trusting a citation. Every source in §6 was sampled

by direct fetch. Four citations were wrong in the first draft and are corrected in place, each

marked CORRECTED at the point of use: the Snyder link served a different report and the subject

and measurement counts attached to it were unsupported; the ANSUR entry attributed a definitions

manual to the wrong authors; the Bammes ISBN resolves to a different publisher and year; and a

ZBrush tutorial was credited with a claim it does not make. Sources that exist but could not be

read (paywall, consent wall, HTTP 403/402) are now labelled as such, and any claim resting on one is

marked at the point of use — §4's landmark list most importantly. No calibration number changed.

The §3 tables were re-checked cell for cell against AGE_PROPORTION_RESEARCH.md and the live JSON

specs and reconcile exactly; the arithmetic in §1.2, §3.4 and §3.5 was recomputed and holds. The

defects were citational, not numerical — which is the good outcome, but it is exactly the class of

defect that makes a proposal-tier document unsafe to build a gate from, so it was worth the pass.

---

1. THE FOUR CANON AUTHORITIES, AND WHAT EACH IS ACTUALLY FOR

Four bodies of work are cited constantly in character-art hiring, review and training. They are not

interchangeable, and the most common mistake in a pipeline is to take a number from the wrong one.

1.1 Zarins, *Anatomy for Sculptors* — the LANDMARK and FORM authority

Uldis Zarins is a Latvian sculptor and the book's origin story — that he is dyslexic and rebuilt

medical anatomy texts as purely visual information usable while sculpting — is the publisher's own

account and is repeated in interviews; it is not verified here (the publisher's site returned 404

on the biography path during the citation audit, §6.0) and nothing in this document rests on it. What

does bind is the book's organising principle, which is checkable on the pages cited below: it is

organised around what the form does on the surface, not around muscle nomenclature.

What it gives that nothing else does:

• A simple-to-complex construction sequence — foundation ("helmethead"/"egghead") → first-level

blockout → second-level blockout → organic form. The head begins as a sphere with the sides sliced

off to establish temples and cheeks; planar structure is committed before any organic

refinement. This is a build order a gate can check by stage, not a style.

• The head division method: cut the top sixth (crown to hairline) off, divide the remainder in

thirds — hairline→brow, brow→nose base, nose base→chin — then split the bottom third in three:

nose→mouth, mouth→upper chin, chin. The head's midpoint lines up with the eyes.

• The braincase-to-face ratio as a growth quantity: roughly 60/40 in the adult, with the

vault finishing early and the face arriving last.

• Proportion charts of the figure and head by age and gender, which is the only artist-side

source in wide studio use that is explicitly age-parameterised.

Their Human Proportions Calculator publishes the age canon it runs on, in Head Units (HU):

newborn 4, infant 5, young child 5.5, child 6, teenager 7, young adult 7.5, idealised adult 8,

elderly 7 — with the explicit caveat that "realistic adult proportions are usually 6 – 7.5 HU" and

that the 8-HU standard is an idealisation from Lysippos, not an anatomical fact.

What it is NOT. It is not a measurement source. Its age ladder is a coarse artist canon in whole

and half units; our measured ladder (§3.1) is finer and disagrees with its "teenager 7" by 0.1 HU,

which is the right size of disagreement and not a conflict.

1.2 Loomis, *Figure Drawing for All It's Worth* — the IDEALISATION authority, and a trap

Loomis presents the ideal figure at 8 heads, and also draws 7.5, 8.5 and 9-head figures; fashion

and entertainment routinely push further still. His torso ladder is one head to the nipples, another

to the navel, another to the crotch, and his shoulder width is 2⅓ heads for men with the hips at

1.5 — all four of those verified in the Proko lesson that teaches his system directly (§6). The

characterisation of the 7.5-head figure as dumpy and old-fashioned is Loomis's own and comes from the

book, which was not read in the citation audit; nothing below depends on it.

**The 2⅓-head shoulder is the single most dangerous number in character art, and our own defect

report is a case of it.** Worked against the measured ladder (§3.2), the bideltoid silhouette width

of a real male is:

Loomis's 2.33 sits **21% above the measured eighteen-year-old and 36% above the measured

fourteen-year-old.** It is an idealisation and it was never anything else. The rejected seven-age

lineup was diagnosed at bideltoid +9.3% at 14 and +12.6% at 18

(AGE_PROPORTION_RESEARCH.md §5, D-3) — the error has exactly the shape of a partially-applied

idealised canon.

The reconciliation, and it is the useful part. The studio heuristic that "a child is about two

heads wide and an adult about three" is *correct* — it is stated in head BREADTHS, not head

heights, and our own measured series confirms it to two decimals: 2.07 at age 7 → 2.61 at 14 →

2.94 at 18 (AGE_PROPORTION_RESEARCH.md §4). Head-height shoulder canons are idealisations;

head-breadth shoulder ratios are measurements. Use the second.

1.3 Bammes, *Die Gestalt des Menschen* — the CONSTRUCTIVE / functional authority

Gottfried Bammes (1920–2007), professor at the Dresden Academy of Fine Arts, wrote what remains the

standard European teaching text. The book was not read in the citation audit (§6.0) — the method

summary here is from catalogue and edition descriptions, and the one claim this document actually

leans on, the ordering claim below, is independently carried by Eaton's verified course structure and

by the verified 80.lv practitioner account. Its method: draw the body from the inside, progressing

from the most general shape to the more specific, with human proportion, body expression, and the

static and dynamic laws of posture and movement as the thematic centres — trunk, then limbs, then head

and features, then the body as a whole entity.

The importable rule is the ordering claim: **the general form is committed before the specific one,

and the skeleton is committed before the muscle.** A pipeline that generates surface detail before

the underlying masses are right cannot be repaired by more detail — which is what a diffusion or

image-to-3D generator does by construction, and therefore the thing our gates must catch.

1.4 Eaton, *Anatomy for Artists* — the CURRENT professional-practice authority

Scott Eaton designed courseware for Pixar's RenderMan and teaches artists from ILM, Pixar, Blizzard

and Disney. His eight-week course sequences skeleton and proportions → torso → back → arms → legs →

hands/feet/neck → variation → animal. Two method points bind on us:

• The core exercise is écorché over reference — drawing the anatomy through a photograph to

visually dissect it. The 3D equivalent is overlaying the anatomy layer on the render and checking

that the surface is explained by something underneath.

• **Bones first, and muscles discussed only in terms of origin, insertion and their effect on the

surface.** Surface form and deep anatomy are treated as one subject, never two.

His bodiesinmotion.photo library is the reference set repeatedly named by working character artists

(e.g. Bjørnar Aarset's 80.lv breakdown, which pairs it with a PureRef board and 2D sketching before

any 3D exploration).

---

2. THE FIVE CONVENTION TRAPS

Every anatomy number in circulation is stated in a convention, and the conventions are not stated.

These five account for essentially all of the apparent disagreement between sources, and four of the

five have already bitten this pipeline.

• Trap 1 — WHICH TOP OF THE HEAD. Anthropometry measures menton→scalp and gets an adult at

8.00 HU. The artistic canon measures to the hair silhouette and gets 7.5. They reconcile

exactly under a constant hair allowance. Half a head unit, and it is the difference between a

figure that reads its age and one that does not. Already diagnosed in

AGE_PROPORTION_RESEARCH.md §2; restated here because every craft source in §1 is in the hair

convention and every measurement source is in the scalp convention.

• Trap 2 — BONE SPAN versus SOFT ENVELOPE. Biacromial is acromion-to-acromion, a skeletal span

taken with a beam calliper. Bideltoid is the maximum horizontal distance across the deltoids, a

soft-tissue envelope. At 14 they differ by 9% (37.03 vs 40.38 cm). A mesh cross-section

measures an envelope; feeding it a skeletal number is a category error, not a tolerance issue.

• Trap 3 — CALLIPER versus TAPE. Calliper breadths compress soft tissue; tape girths do not. A

breadth and a girth taken over the same section do not describe the same surface, and a mesh

reproduces the *uncompressed* one.

• Trap 4 — RESPIRATORY PHASE. Chest breadth is taken at maximum inspiration and chest depth at

quiet respiration in the standard protocols. Two numbers from one table are not from one instant

of one breath.

• Trap 5 — THE ELLIPSE SHORTCUT IS FORBIDDEN. Deriving chest depth from breadth and

circumference through an ellipse is wrong by about 30%. A thorax is not an ellipse.

Traps 3–5 are already recorded as binding in AGE_PROPORTION_RESEARCH.md §9 and are repeated here

because anatomy_bar.py is a new consumer that has not yet inherited them.

---

3. THE AGE-14 MALE CALIBRATION TABLES

Everything in this section is for the protagonist rung A14, male, at the ladder stature

166.72 cm. Sources are keyed: [SNY] Snyder et al. 1977 (UM-HSRI-77-17, subject-level file,

recomputed — not the 1975 UM-HSRI-BI-75-5 child report, §6.0); [NH3] the NHANES III LMS

references, which AGE_PROPORTION_RESEARCH.md §3 keys to Hawkes et al. (2020) sitting-height and

leg-length LMS plus the NHANES III breadth tables; [KRK] Klimek-Piotrowska et al. 2015 (Cracow

adolescents, PLOS ONE); [A4S] Anatomy for Sculptors; [LOO] Loomis (prose only — no row is keyed

to it, by design). Rows marked DERIVED are computed here from cited rows and say so.

**Every [SNY] and [NH3] fraction in §3.1 and §3.2 was re-checked against

AGE_PROPORTION_RESEARCH.md in the 2026-08-08 audit and matches it cell for cell** (crotch height

0.4901, biacromial 0.2221, biiliac 0.1559, head breadth 0.0928, bideltoid 0.2422, head height

20.82 cm and 7.75/7.10 HU at 14). The centimetre values here are those fractions times the ladder

stature 166.72 cm, which is why the head-breadth cell reads 15.47 cm where the sibling document's

raw Snyder mean reads 14.90 cm at a smaller Snyder-era stature; the ratio R-2 depends on,

0.2422 ÷ 0.0928 = 2.610, is scale-free and is unaffected. The [AN] key from the first draft has

been retired — the ANSUR attribution it pointed at was wrong (§6.0) and no row used it.

3.1 The proportion ladder — head units and vertical landmarks

Three artist rules checked against that table, because a gate should know how far each is off before

it trusts one:

• "Legs are three-and-a-half to four heads." At 14 the measured value is 3.48 HU — just under

the bottom of the artist range, and correctly so, because the adolescent long-legged phase is

measured against a head that is nearly adult-sized. The rule is usable as a floor, not as a target.

• "The hand is as long as the face." Face length (chin→hairline) is a constant 0.82–0.83 of head

height at every age; at 14 that is 17.18 cm against a hand of 18.41 cm. The hand is **7% longer

than the face at 14**, and still 5% longer at 18. The rule is approximately true for adults and

systematically wrong in the adolescent direction — which is the same phenomenon as "hands and feet

arrive before stature."

• "The foot is head-sized." Foot 26.26 cm against a hair-convention head of 23.48 cm: the foot is

12% longer. Against the scalp head it is 26% longer. Do not use this rule as a check.

3.2 Breadths — the silhouette widths

The load-bearing ratio for a review gate is bideltoid ÷ head breadth, not shoulder-over-stature.

Shoulder-over-stature changes only 5.7% across eleven years; head-breadth-over-shoulder changes 29%.

When a young figure reads "shoulders too broad," the defect is usually a head that is too small to

measure the shoulders against — there is no other reference in the silhouette. That is the mechanism

behind our own D-1/D-3 pair, where the head was 4.5–7.2% too small at every rung *and* the shoulders

read broad.

3.3 Girths — the anti-gaunt row set

These are the numbers a gaunt mesh fails and a proportion check does not look at. Both live specs are

shown because they disagree, and the disagreement is itself a finding (§7 and GAP-1).

Independent cross-check from a measured adolescent population [KRK], n = 121 boys at age 14:

height 165.7 ± 8.5 cm, mass 55.5 ± 11.1 kg, BMI 20.1 ± 3.0, waist 73.6 ± 9.7 cm, hip

87.9 ± 8.4 cm, waist-to-height 0.44 ± 0.05, waist-to-hip 0.91 ± 0.05, body fat 20.0 ± 11.6%. Our

ladder figure (BMI 19.6, waist-to-height 0.413, waist-to-hip 0.824) sits inside that distribution on

BMI and slightly lean on waist — which is intended, and which is exactly why a leanness knob with no

floor under it is dangerous (GAP-4).

3.4 Mass and volume — the cheapest gaunt detector we are not running

A watertight mesh has a measurable volume, and volume is the one quantity a gaunt figure cannot

fake. Whole-body density at 20% fat is 1.05 g/cm³ by the Siri relation (D = 495 / (%BF + 450)).

Proposed bar. A closed A14 mesh whose volume implies BMI below mean − 1 SD (17.1) is a

WARNING; below mean − 2 SD (14.1) it is a FAILURE, named gaunt, regardless of every proportion

check passing. Proportion is a ratio test and is *scale-free* — it is structurally incapable of

seeing that a correctly-proportioned figure has been shrink-wrapped.

3.5 The fat mantle — millimetres, at named sites

This is the number that converts "reads gaunt" from taste into measurement. A skinfold is a double

fold of skin plus subcutaneous fat, so the tissue depth over the site is half the caliper reading.

From [KRK], boys aged 14, n = 121:

The craft rule this licenses, stated so a gate can hold it. At age 14 there are **five to seven

millimetres of skin and fat** over the scapula and the iliac crest. Those landmarks are therefore

soft ridges and plane changes, never hard edges, and never separated silhouette forms. A mesh in

which the scapular border, the iliac crest, the ribs or the vertebral spinous processes below C7 read

as discrete raised edges is not "more anatomical" — it is a flayed figure, an écorché wearing no

mantle, and it is wrong by measurement at those five sites.

---

4. THE LANDMARK CHECKLIST STUDIO ARTISTS REVIEW AGAINST

Citation status, stated before the list. This list was drawn from a Creative Bloq ZBrush anatomy

article whose body was not retrievable to automated fetch on two attempts in the 2026-08-08 audit

(§6.0) — the page and title are real, the text was not readable. The list is therefore carried as

standard descriptive anatomy, independently checkable in any anatomy text, not as a quotation from

a verified source: the sternum, costal cartilage, iliac crest, vertebrae, anterior superior iliac

spine, curve of the tibia, clavicle, scapula, acromion process, zygomatic bone, mandible and temporal

line are subcutaneous or near-subcutaneous by anatomical fact, and the scapula's **spine and medial

(inner) border** are the two parts of it lying directly under skin while the rest is muscle-covered.

Nothing here is an artistic convention that could be wrong; it is bone position. Before R-6 ships,

re-source this list to a readable citation — a checklist that cannot name its authority is the same

defect as a band typed from taste.

Split for our A14 target, that becomes a three-column gate. This is the table anatomy_bar.py should

carry as its landmark register.

4.1 MUST READ — present and locatable at age 14

• Clavicle, full length, with its S-curve and the sternal and acromial ends both findable.
• Acromion process — the corner of the shoulder, and the top of the biacromial span.
• Spine of the scapula and the medial border — as soft ridges under a 5.3 mm mantle.
• Sternum / jugular notch (suprasternale) — at 0.8115 of stature, the neck-to-chest transition.
• Iliac crest and ASIS — as a plane change under a 5.5 mm mantle, felt more than seen.
• Patella, and the curve of the tibia (the anterior crest) down the front of the shank — the

tibia is genuinely subcutaneous and is *supposed* to read.

• Medial and lateral malleoli, asymmetric in height (medial higher).
• Olecranon at the elbow; ulnar styloid at the wrist.
• Mandible, zygomatic arch, temporal line on the head.
• Cricoid/thyroid prominence — present at 14, understated; the larynx is arriving.

4.2 MUST NOT READ — a positive finding if present at this age and this adiposity

• Individual ribs as separate raised bars anywhere on the front or the side of the thorax.
• Vertebral spinous processes below C7 as a beaded chain.
• Serratus digitations as discrete fingers.
• The iliac crest as a hard silhouette edge rather than a plane change.
• Deltoid, biceps and triceps separated by hard troughs — at surface_definition_index 0.3 the

definition is emerging and subtle, not carved.

• Tendons of the forearm and the dorsum of the hand as raised cords.
• Hollowed temples, hollowed cheeks under the zygomatic, sunken supraclavicular fossae. The

supraclavicular fossa in particular: a deep pit there is the classic single tell of a gaunt figure

and is the first thing to check in a bust render.

4.3 AGE-SPECIFIC SHAPE CUES — what proportion numbers cannot fix

Head units are nearly exhausted by 16: there is about 0.4 HU of total budget between 14 and 18,

so a ladder that varies only stature and head scale is mathematically forced to make 14, 16 and 18

one figure. The separation has to come from shape. In our own ranked order:

1. Head held near-constant while the body scales.

2. Cranial vault versus face — the vault finishes early, the jaw and brow arrive last.

3. Leg-to-trunk ratio and its inversion — the cormic index falls to a minimum near 13 and rises

again, so a thirteen-year-old's legs are proportionally longer than a grown man's. A monotone

progression cannot express this and is the most damaging single artefact.

4. The shoulder yoke — narrow, sloped, no deltoid cap, no trapezius yoke before puberty.

5. The neck — short, thin, head low and forward, undeveloped trapezius.

Two corrections that matter for a fourteen-year-old specifically: the waist taper should be built as

wide shoulders, not as a narrow waist; and lumbar lordosis *increases* with age (30° at 2–4 →

44° at 17–20), which is the opposite of the usual instinct.

---

5. THE REVIEW PROTOCOL — the gates a AAA character passes

Provenance, stated plainly. The three-gate protocol below is taken from one source — the

*Stylized 3D Characters: Art Direction Principles for AAA-Tier Production* playbook (§6), a

practitioner blog post, not a first-party studio document. It is the most explicit written statement

of the protocol found, and the two GDC character-pipeline talks in §6 are cited as pipeline

references only — neither was verified to contain this three-gate/five-pose protocol. Adopt it as

a well-formed proposal we are choosing, not as an industry standard we are quoting.

Three hard gates, each with a comparison rather than an opinion.

• Blockout gate — the silhouette test. The character must be identifiable as a **black silhouette

alone** at three distances: hero (~5 m), gameplay-typical (~15 m) and recognition-threshold

(50–100 m, camera-dependent). The source specifies the five canonical poses — neutral idle, attack,

traversal, hit reaction, victory — at the recognition threshold; running all five at all three

distances is our own extension, and it is named as ours. Failure sends the asset back to blockout;

it is not fixed downstream.

• High-poly gate — silhouette comparison against the concept's flat silhouette study, with the

concept overlaid at the same camera.

• Final-bake gate — side-by-side hero shots at matched lighting against the approved concept,

and a hero comparison render with the rest of the cast under one light.

Two further practices from working character artists that translate directly into gate steps:

• Work small to fix big problems, work big to fix small problems (Marshall Vandruff, quoted by

Aarset). Zoom out to judge primary form; zoom in only for tertiary. A gate that only ever renders

close-ups cannot see a proportion error.

• Primary → secondary → tertiary is a hard ordering. If the primary forms are wrong, the

secondary and tertiary forms will look bad no matter how well they are made — which is precisely

the failure mode of a generator that produces plausible microdetail on wrong masses.

---

6. SOURCES

6.0 Citation audit — 2026-08-08

Every source below was sampled by direct fetch on 2026-08-08. Each entry now carries its

verification state, and four defects found in the first draft were corrected in place:

the Snyder link pointed at a different report, the ANSUR entry attributed a definitions manual to

the wrong authors, the Bammes ISBN resolved to a different publisher, and one practitioner

article was credited with a claim it does not make. States used below:

• VERIFIED — fetched, and the page says what this document says it says.
• EXISTS, BODY NOT READ — the page is real and correctly titled, but its text was not retrievable

to automated fetch (paywall, consent wall, HTTP 403/402). Any claim drawn from it is marked.

• CORRECTED — the first draft misstated the citation; the corrected form is below.

Books and manuals

• Zarins, Uldis, with Kondrats, Sandis. *Anatomy for Sculptors: Understanding the Human Figure.*

Exonicus / Anatomy Next (trading as Anatomy for Sculptors), first edition 2014

(ISBN 9780990341109 — Open Library records that ISBN under Exonicus, 232 pp; 2020 printing

9781735039022). Includes proportion charts of the figure and head by age and gender.

EXISTS (ISBN resolved via Open Library; the book itself was not read here — the two publisher

pages below are what this document actually quotes). https://anatomy4sculptors.com/

• Anatomy for Sculptors — "Forming the Head: Simple to Complex" (head construction sequence

"foundation → 1st-level blockout → 2nd-level blockout → organic forms"; helmethead and egghead

foundations; "divide the rest into thirds: hairline to brow, brow to the bottom of the nose, and

nose to chin"; "Adults have more balanced skulls – about 60/40 between the braincase and face").

VERIFIED. https://anatomy4sculptors.com/blog/forming-the-head-simple-to-complex/

• Anatomy for Sculptors — "About Human Proportions Calculator" (the 8-band HU-by-age canon: newborn 4,

infant 5, young child 5.5, child 6, teenager 7, young adult 7.5, adult 8, elderly 7; "Realistic

adult proportions are usually 6 – 7.5 Hu"; the 8-HU standard attributed to Lysippos, 4th century BC).

VERIFIED — every value in §1.1 matches.

https://anatomy4sculptors.com/blog/about-human-proportions-calculator/ ·

tool: https://hpc.anatomy4sculptors.com/

• Loomis, Andrew. *Figure Drawing for All It's Worth.* Viking, 1943 (Titan reissue 2011). The 8-head

ideal, the 7.5/8.5/9-head variants, the 2⅓-head male shoulder. The book itself was not read here

— the 2⅓ figure this document argues against is taken from the Proko lesson below, which teaches

Loomis's system directly. Author reference: https://en.wikipedia.org/wiki/Andrew_Loomis

• Proko — "Human Proportions: Idealistic Figures" (Loomis's system taught; "Loomis makes the shoulders

a bit wider at 2 ⅓ heads. He keeps the hips at 1.5 heads."; "One head down to the nipples, another

to the navel, and another to the crotch."). VERIFIED — this is the actual source of the 2⅓ figure.

https://www.proko.com/course-lesson/human-proportions-idealistic-figures/

• Bammes, Gottfried (1920–2007). *Die Gestalt des Menschen: Lehr- und Handbuch der Künstleranatomie.*

Multiple editions across three publishers; **ISBN 9783862300013 resolves to the Englisch Verlag GmbH

edition, September 2010, 514 pp (Open Library) — CORRECTED**, the first draft attributed that

ISBN to "Ravensburger / Verlag Christophorus", which it is not. Dresden Academy method; drawing the

body "from the inside"; general-to-specific ordering; proportion, expression, and the static and

dynamic laws of posture and movement. The book was not read here; the method summary in §1.3 is

from the catalogue descriptions and from the English translation's stated scope, and the ordering

claim §1.3 actually leans on is independently carried by Eaton (verified) and by the 80.lv

practitioner source (verified). Author: https://en.wikipedia.org/wiki/Gottfried_Bammes ·

works (28 titles incl. *Die Gestalt des Menschen* 1997 and 2002 editions, and the English

*The Artist's Guide to Human Anatomy*): https://openlibrary.org/authors/OL597973A/Gottfried_Bammes

• 3dtotal Publishing (ed.), with Chris Legaspi and Mario Anger as featured contributing artists.

*Anatomy for 3D Artists: The Essential Guide for CG Professionals.* 3dtotal Publishing

(ISBN 9781909414242; 288 pp; now out of print). CORRECTED — the first draft credited Legaspi as

author; the publisher's own page describes him as one of the "established artists" featured, and

Open Library credits "3dtotal Publishing Staff and Chris Legaspi". Covers 2D reference → sculpting →

topology. EXISTS, BODY NOT READ — the "labelled male reference figure in front and side" detail

claimed in the first draft was not confirmed and has been dropped.

https://store.3dtotal.com/products/anatomy-for-3d-artists

• 3dtotal Publishing (illus. Charlie Pickard; introduction by Jahirul Amin). *Anatomy for Artists: A

Visual Guide to the Human Form.* ISBN 9781912843107, hardback, 304 pp. "over 240 specially

commissioned photographs" with "musculature, planar, and contour diagrams drawn by figurative fine

artist and illustrator Charlie Pickard." VERIFIED.

https://store.3dtotal.com/products/anatomy-for-artists

• Winter, David A. *Biomechanics and Motor Control of Human Movement*, 4th ed. Wiley, 2009 —

reproduces the Drillis & Contini segment-length-as-fraction-of-stature table. **EXISTS, BODY NOT

READ** (publisher returned HTTP 402 to automated fetch). No number in this document is taken from it.

https://onlinelibrary.wiley.com/doi/book/10.1002/9780470549148

Courses, talks and practitioner material

• Eaton, Scott. *Anatomy for Artists* (online course; "DURATION: 8 weeks + 4 months additional access";

eight weekly lectures running proportions/skeleton → chest, shoulders, abdomen → back → arms →

pelvis and legs → hands, feet, neck → gender/age/weight variation → animal anatomy; écorché

assignments in which artists "visually dissect the work of Michelangelo, old master drawings,

sculptures from the Louvre"; "The entire first lecture is dedicated to proportions, bones, and the

skeleton… each weekly lecture talks about the bones in relation to origin and insertions of the

muscles and how they influence the surface"; Eaton "designed and authored Pixar's RenderMan

Courseware", client studios listed include ILM, Epic Games, Blizzard, Pixar and Disney).

VERIFIED — every claim in §1.4 matches. https://www.scott-eaton.com/anatomy-for-artists

• Eaton, Scott. *Bodies in Motion* reference library. https://www.bodiesinmotion.photo/
• 80 Level — "Studying Organic Character Art: Anatomy & Skin" (Bjørnar Aarset, 3D character artist,

Oslo: "This is pretty standard stuff: Primary>Secondary>Tertiary."; "A quote I heard Marshall

Vandruff mention was, work small to fix big problems and work big to fix small problems.";

"usually from bodiesinmotion.photo by Scott Eaton"; "a big reference canvas in PureRef").

VERIFIED — all four attributions in §1.4 and §5 confirmed verbatim.

https://80.lv/articles/001agt-studying-organic-character-art-anatomy-skin

• 80 Level — "80 Level Digest: Materials on Human Anatomy for 3D Artists" (the resource roster

working artists actually name). NOT SAMPLED in the 2026-08-08 audit.

https://80.lv/articles/80-level-digest-materials-on-human-anatomy-for-3d-artists

• Creative Bloq — "Sculpt realistic anatomy in ZBrush." EXISTS, BODY NOT READ — the page and its

title are confirmed, but the article body was not retrievable to automated fetch on two attempts

(the fetch returns navigation chrome and truncates before the body). The landmark list and the

scapula claim in §4 are therefore unconfirmed from this citation. They are not thereby wrong —

the subcutaneous status of the scapular spine and medial border is standard descriptive anatomy and

is independently carried by the Eaton course structure — but §4 is marked accordingly and the

reconciliation is owed. https://www.creativebloq.com/how-to/sculpt-realistic-anatomy-in-zbrush

• Creative Bloq — "10 anatomy tips for 3D artists." EXISTS, BODY NOT READ (same truncation).

https://www.creativebloq.com/10-anatomy-tips-3d-artists-3113091

• "Stylized 3D Characters: Art Direction Principles for AAA-Tier Production" — CORRECTED title

(the first draft called it "Stylized 3D Characters Done Right: The Production-Grade Art Direction

Playbook", which is not the page's title). A practitioner blog post, not a first-party studio

document. VERIFIED for: "blockout review with the concept overlaid at the same camera; high-poly

review with silhouette comparison against the concept's flat silhouette study; final-bake review

with side-by-side hero shots at matched lighting"; "Render the character at three test distances:

hero-shot close (5 meters), gameplay-typical (15 meters), and recognition-threshold (50–100

meters)"; "Test the character in five canonical poses — neutral idle, attack pose, traversal pose,

hit reaction, and victory pose — at the recognition threshold"; proportion ratios (head-to-body,

limb-to-torso, shoulder breadth) locked in the art bible before 3D production. **Note the last

correction**: the five poses are specified at the recognition threshold, not across all three

distances — §5 now says so. https://nastyrodent.com/stylized-3d-characters-art-direction-principles/

• GDC Vault — "Uncharted 2 Character Pipeline: An In-depth Look at the Creation of U2's Characters,"

Richard Diamant and Judd Simantov, Naughty Dog, GDC 2010, Visual Arts track. **VERIFIED to exist,

CONTENT NOT VIEWED** — cited as a character-pipeline reference only; no claim in this document is

sourced to it. https://gdcvault.com/play/1012552/Uncharted-2-Character-Pipeline-An

• GDC Vault — "Building Champions: Character Art Pipeline for 'Rumbleverse'," Steven Kosanovich, Iron

Galaxy Studios, GDC 2023, Visual Arts track. VERIFIED to exist, CONTENT NOT VIEWED — same

standing. https://www.gdcvault.com/play/1028733/Building-Champions-Character-Art-Pipeline

• Novedge — "Enhancing Anatomy Sculpting Skills in ZBrush" (8 June 2024). CORRECTED — the page

supports only the layer-discipline claim: "Utilize layers in ZBrush to experiment with different

levels of muscular definition without losing the ability to step back to a previous state." The

first draft additionally credited it with "definition varies with pose, age, body type and tension —

relaxed muscle transitions softly, flexed muscle firms and shows tendon," which **the page does not

say**; that claim has been removed. (The pose/tension variation itself is real and is carried by

Eaton's week-7 gender/age/weight-variation lecture, verified above.)

https://novedge.com/blogs/design-news/zbrush-tip-enhancing-anatomy-sculpting-skills-in-zbrush-tips-and-techniques-for-realistic-character-models

• Envato Tuts+ — "Human Anatomy Fundamentals: Drawing Different Ages," Joumana Medlej (age HU ladder:

child "5.5 heads in early childhood, to 6 heads between 7 and 9"; adolescent "Height= 6.5 to 7

heads"; young adult 7.5; "In males, shoulders broaden."; "The feet, arms, legs and hands may grow

faster than the rest of the body, leading to the gangly look"). VERIFIED.

https://design.tutsplus.com/tutorials/human-anatomy-fundamentals-drawing-different-ages--cms-21905

• Winged Canvas — "How to Draw Human Proportions: Young and Old" ("A male's ribcage is approximately

2 times the volume of his head."; "your fingertips end at approximately a third of your thigh";

"Your hand is typically the size of your face (minus the hairline)!"; "Feet on adults are

approximately the size of their head."). VERIFIED.

https://www.wingedcanvas.com/single-post/how-to-draw-human-proportions-young-old

• Wikipedia — "Body proportions" ("An average person is generally 7-and-a-half heads tall"; ideal 8;

"A heroic figure… is eight-and-a-half heads tall"; legs "three-and-a-half to four heads long"; arms

"about three heads long"; "hands are as long as the face"; shoulder-to-hip "average ratio for women

is 1:1.03, for men it is 1:1.18"; "The half-way mark is a line between the greater trochanters").

VERIFIED. https://en.wikipedia.org/wiki/Body_proportions

Anthropometry

• Snyder, R.G., Schneider, L.W., Owings, C.L., Reynolds, H.M., Golomb, D.H., Schork, M.A. (1977).

*Anthropometry of Infants, Children and Youths to Age 18 for Product Safety Design.* UM-HSRI-77-17,

Highway Safety Research Institute (now UMTRI). CORRECTED, twice. (a) The first draft said

"4,127 subjects, 87 measurements." Neither figure is supported. The owning document,

AGE_PROPORTION_RESEARCH.md §2, records the UMTRI subject-level file as **3,901 subjects × 123

columns, male subsample n = 1,978**, recomputed from raw and validated against the report's own

published male tables at five age bands — those are the numbers, and they are the ones every [SNY]

row in §3 rests on. (b) The first draft's second link,

mreed.umtri.umich.edu/…/child/Snyder_1975_Child.pdf, is a different report — it downloads as

Snyder, Spencer, Owings & Schneider (May 1975), *Physical Characteristics of Children as Related to

Death and Injury for Consumer Product Safety Design*, UM-HSRI-BI-75-5, 4,027 infants and

children newborn to 12 years, 41 measurements (verified by download and text extraction on

2026-08-08). It cannot source an age-14 or age-18 row and is no longer cited as though it could;

it is listed below on its own terms. The head-height definition (report p. 168, menton→vertex

perpendicular to the Frankfort plane) is carried from AGE_PROPORTION_RESEARCH.md §2 and was **not

re-verified here** — the Deep Blue landing page returns HTTP 403 to automated fetch.

https://deepblue.lib.umich.edu/handle/2027.42/684

• Snyder, R.G., Spencer, M.L., Owings, C.L., Schneider, L.W. (1975). *Physical Characteristics of

Children as Related to Death and Injury for Consumer Product Safety Design.* UM-HSRI-BI-75-5,

Highway Safety Research Institute, contract FDA 72-70. 4,027 subjects newborn–12 years, 41

measurements (12 of them "applied measurements which have not been previously available").

VERIFIED by download. Listed for completeness and for the child rungs; not the source of the

A14/A18 rows. https://mreed.umtri.umich.edu/mreed/downloads/anthro/child/Snyder_1975_Child.pdf

• Klimek-Piotrowska, W., Koziej, M., Hołda, M.K., et al. (2015). "Anthropometry and Body Composition

of Adolescents in Cracow, Poland." *PLOS ONE.* DOI 10.1371/journal.pone.0122274. Age-14 boys

n = 121: height 165.7 ± 8.5 cm, mass 55.5 ± 11.1 kg, BMI 20.1 ± 3.0, waist 73.6 ± 9.7 cm, hip

87.9 ± 8.4 cm, W/Ht 0.44 ± 0.05, W/H 0.91 ± 0.05; skinfolds biceps 7.6 ± 5.2, triceps 13.1 ± 7.8,

subscapular 10.6 ± 7.3, suprailiac 10.9 ± 8.1 mm; body fat 20.0 ± 11.6%. **VERIFIED — every one of

the thirteen values above was read back off the article and matches this document cell for cell.

This is the load-bearing citation for §3.3, §3.4, §3.5, R-5 and R-12, and it holds.**

https://pmc.ncbi.nlm.nih.gov/articles/PMC4376782/

• NCHS Vital and Health Statistics, Series 11 No. 249 (April 2009). *Anthropometric Reference Data

for Children and Adults: United States, 1988–1994* — the published biacromial-breadth table for

ages 3–19 by sex (mean, SE, selected percentiles). https://www.cdc.gov/nchs/data/series/sr_11/sr11_249.pdf

*(HTTP 403 to automated fetch from here on every CDC host tried — the same block

AGE_PROPORTION_RESEARCH.md §6 recorded. Cited as the reference standard, not read here.)*

• NCHS Vital and Health Statistics, Series 11 No. 252. *Anthropometric Reference Data for Children

and Adults: United States, 2007–2010.* https://www.cdc.gov/nchs/data/series/sr_11/sr11_252.pdf

• *Anthropometric Data Analysis Sets (ADAS) Manual.* Human Systems Information Analysis Center,

AFRL/HEC/HSIAC, Wright-Patterson AFB, OH. CORRECTED — the first draft attributed this PDF to

"Gordon, C.C., et al. (2014), *2012 Anthropometric Survey of U.S. Army Personnel (ANSUR II)*."

That is the wrong document: the file's own title page reads "ANTHROPOMETRIC DATA ANALYSIS SETS

MANUAL, PREPARED BY: Human Systems Information Analysis Center AFRL/HEC/HSIAC." ANSUR II is a real

and separate report; it is simply not what this URL serves. **The two definitions Trap 2 rests on

are VERIFIED verbatim in the correct document** (downloaded and text-extracted 2026-08-08):

biacromial breadth — "The distance between the right and left acromion landmarks at the tips of the

shoulders is measured with a beam caliper. The subject sits erect."; bideltoid breadth — "The

maximum horizontal distance between the lateral margins of the upper arms on the deltoid muscles is

measured with a beam caliper." Trap 2 is therefore sound; only its citation was wrong.

https://mreed.umtri.umich.edu/mreed/downloads/anthro/ansur/ADAS-Dimension_Definitions.pdf

• Rollins, J.D., Collins, J.S., Holden, K.R. (2010). "United States Head Circumference Growth

Reference Charts: Birth to 21 Years." *The Journal of Pediatrics.* NOT READ — publisher returned

HTTP 403 to automated fetch on 2026-08-08. No number in this document is taken from it; it is

carried as an onward cross-check. https://www.jpeds.com/article/S0022-3476(10)00020-X/pdf

• Farkas, L.G. (1994). *Anthropometry of the Head and Face*, 2nd ed. Raven Press — the independent

craniofacial cross-check. NOT READ (lending-restricted; GAP-9 records this and it is the

outstanding work item, not a claim source). https://archive.org/details/anthropometryofh0000unse ·

Farkas, Posnick, Hreczko (1992), "Anthropometric Growth Study of the Head," *Cleft Palate-Craniofacial

Journal.* https://journals.sagepub.com/doi/abs/10.1597/1545-1569_1992_029_0303_agsoth_2.3.co_2

• Drillis, R., Contini, R. (1966). *Body Segment Parameters.* Report 1166-03, NYU School of

Engineering — the segment-length-as-fraction-of-stature constants. **The report itself was not read

here, and the two Penn State OPEN Design Lab pages below do NOT carry either the report's

bibliographic detail or its fraction table** — they carry only the method and the attribution

("Drillis and Contini were among the first to publish mathematical ratios of many body dimensions to

stature (Drillis and Contini, 1966)"). Verified to that extent and no further. No [SNY], [NH3]

or [KRK] row in §3 depends on Drillis & Contini.

https://www.openlab.psu.edu/design-tools-proportionality-constants/ ·

https://www.openlab.psu.edu/project/proportionality-constants/

• Siri, W.E. (1961). "Body composition from fluid spaces and density: analysis of methods." The

%BF ↔ density relation D = 495 / (%BF + 450) used in §3.4. NOT RE-VERIFIED here — the relation

is standard and the arithmetic in §3.4 is shown, so it is checkable without the paper.

---

7. ACTIONABLE RULES — mapped to our stack

Each rule names the file that owns it. Rules R-1 through R-8 are the calibration contract for

harness/asset_factory/anatomy_bar.py; R-9 through R-13 bind the surrounding lane.

R-1 · The bar declares its convention on every site, or it refuses

anatomy_bar.py carries a convention field per measured site with exactly one of

scalp | hair | bone_span | soft_envelope | calliper_breadth | tape_girth, and refuses to compare

two sites whose conventions differ. This is Traps 1–3 made structural rather than remembered. The

existing precedent is check_anatomy_conformance.py's angle_convention field — same pattern, same

reason.

R-2 · Shoulder width is scored in HEAD BREADTHS, against the bideltoid envelope

Target at A14: bideltoid ÷ head breadth = 2.61, from bideltoid 40.38 cm and

head breadth 15.47 cm. The bar must not score shoulder-over-stature as a primary (it moves only

5.7% across eleven years and cannot discriminate), and must never apply a head-height shoulder

canon — Loomis's 2⅓ is 36% above the measured fourteen-year-old (§1.2). Ladder for the other rungs:

2.07 / 2.24 / 2.33 / 2.46 / 2.61 / 2.85 / 2.94 at ages 7 / 9 / 11 / 12.5 / 14 / 16 / 18.

R-3 · A head-size lens runs BEFORE any shoulder lens, and gates it

Given §3.2, a "shoulders too broad" finding is not trustworthy until the head is confirmed. The bar

orders head_units → head_breadth → bideltoid, and if the head lens is outside tolerance the

shoulder lens reports INCONCLUSIVE_HEAD_REFERENCE instead of a broadness finding. Our own D-1/D-3

pair is the fixture: head 4.5–7.2% small at every rung *and* shoulders reading broad.

R-4 · Volume is measured, and gaunt is a named failure (§3.4)

On a closed mesh, mass = volume_L × 1.05, BMI = mass / 1.6672². WARNING below BMI 17.1, FAILURE

named gaunt below 14.1, at A14. This is the one test that a scale-free proportion check cannot

substitute for, and it is cheap. It belongs in anatomy_bar.py and its result should also be

surfaced by harness/asset_factory/thousand_point_eval.py as a ninth lens, since that is where the

lens battery lives.

R-5 · The fat-mantle rule is a measurement, not taste (§3.5)

At five named sites the bar asserts a minimum surface softness consistent with the measured tissue

depth: triceps 6.6 mm, suprailiac 5.5 mm, subscapular 5.3 mm, biceps 3.8 mm. Operationally, on the

region rings thousand_point_eval.py already casts, the **local radial curvature at those sites must

not exceed that of a fillet of the stated radius** — a hard edge at the scapular border or the iliac

crest is a positive flayed finding. This is the MUST NOT READ list in §4.2 given a number.

R-6 · The landmark register from §4 ships as data, not prose

anatomy_bar.py carries the §4.1 MUST READ and §4.2 MUST NOT READ lists as a table keyed by

region, matching the region partition BODY_AGE_14.json :: regions already uses (R01_FOOT, …), so

a finding names a region the rest of the pipeline already knows. MUST NOT READ items are scored on

the same rings as R-5.

R-7 · Every band is DERIVED and every band is declared — but never from two disagreeing specs

thousand_point_eval.py already holds the right principle ("an acceptance band typed from taste is

the same defect as a claim without a citation"). The implementation it chose — the band is the

disagreement between BODY_AGE_NN.json and BODY_AGE_NN_RESEARCH.json — is the defect described in

GAP-1. anatomy_bar.py must take one spec as target and derive its band from a population SD

([KRK], [SNY], [NH3] all publish one), never from the spread between two of our own files.

R-8 · Vacuity control, inherited

A lens that scores zero points reports VACUOUS and VACUOUS is a failure. A site the bar cannot locate

REFUSES by name. This is already law in check_anatomy_conformance.py and thousand_point_eval.py

and anatomy_bar.py inherits it verbatim — an instrument that cannot find anything must never report

a clean bill of health identical to a real one.

R-9 · The silhouette gate gets the three distances and the five poses (§5)

build/3d/best_a14/silhouette_overlay.py currently compares one framing (verified: its --spec

default is BODY_AGE_14.json, line 75). Extend it to the protocol in §5: black-silhouette recognition

at hero / gameplay / recognition-threshold framings, with the five poses — neutral idle, attack,

traversal, hit reaction, victory — at least at the recognition threshold, which is what the source

specifies; running all five at all three framings is our extension and is cheap. A blockout failure

returns to blockout; it is never repaired downstream. **Provenance caveat: §5 has one source and it is

a practitioner blog, not a studio document.** The rule is worth adopting on its merits; it is not

being adopted because an authority mandates it.

R-10 · Primary before secondary before tertiary, enforced by stage

Per Bammes and Eaton and every practitioner source in §6: if the primary masses are wrong the

tertiary detail cannot save them. The lane's stage DAG should refuse to run refinement passes

(seedvr2_refine.py, normals_tooth.py, texture bake) on a mesh that has not cleared the

anatomy_bar.py primary-form verdict. This is the generator-specific failure mode — plausible

microdetail on wrong masses — and it is the shape of the defect Josh caught.

R-11 · The age-shape cues are built, not scaled (§4.3)

From 14 to 18 there is ~0.4 HU of total budget. harness/asset_factory/aging_ladder.py,

age_keystones.py and age_anchors.py must express the cormic-index inversion (a non-monotone

crotch-height fraction, peaking near 12.5) and the shoulder yoke as shape, not as scale. A

monotone leg-length series is a positive finding, not a tuning choice.

R-12 · The leanness knobs get a floor

BODY_AGE_14.json :: training_clock.adiposity = -0.45 and subcutaneous_fat_scalar = 0.8995 are

unbounded below by anything measured. Clamp them so the *implied* body fat cannot fall below

[KRK] mean − 1 SD, and have anatomy_bar.py re-derive the implied value rather than trusting the

knob. The prompt_physique string in that same block is already doing the right thing in words —

"the whole torso wrapped in even healthy flesh so the surface over the ribs and the sides is smooth

and unbroken" — and R-5 is that sentence with millimetres attached.

R-13 · Reference is sculpture, not photographs of minors

Standing, and restated because this document's craft half is the one that would otherwise reach for

photo reference: use museum sculpture. The Spinario (Capitoline) is the pre/early-pubertal read —

thin neck, no deltoid cap, boxy torso, cylindrical limbs; the Getty Victorious Youth has the yoke

and cap fully present. The gap between those two bronzes is our 12.5→18 delta and five of the ranked

shape cues are visible in that single comparison. Use anthropometric numbers; do not retrieve the

measurement-position plates that accompany them.

---

8. WHAT WE ADOPT

• Adopt the Zarins construction order and the Zarins head divisions as the head-lens structure

(facial thirds; lower third in thirds; midpoint at the eyes; 60/40 vault-to-face as an adult

endpoint with the vault-first growth gradient toward it). It is the only widely-used artist system

that is explicitly age-parameterised.

• Adopt Bammes's and Eaton's ordering as a pipeline law: general form before specific, skeleton

before muscle, primary before tertiary — enforced as R-10, a stage gate, because it is the one

discipline a generative pipeline structurally lacks.

• Adopt the head-BREADTH shoulder ratio (R-2) and reject every head-HEIGHT shoulder canon.

Loomis's 2⅓ is an idealisation that is 21% above the measured adult and 36% above our A14 target,

and the measured lineup defect (+9.3% at 14, +12.6% at 18) has its exact signature.

• Adopt the volumetric gaunt test (R-4). It is the cheapest high-value gate available to us, it

catches the exact failure that shipped, and a ratio-based proportion check is structurally

incapable of it.

• Adopt the fat-mantle millimetres (R-5) and the MUST-READ / MUST-NOT-READ register (R-6). This

converts "reads gaunt" and "flayed" into measured findings at named sites, which is what the

standing evidence discipline requires of any claim.

• Adopt the three production review gates (R-9) — blockout silhouette at three distances, the five

poses at least at the recognition threshold, high-poly silhouette against the concept, final-bake

side-by-side at matched lighting. **Adopted as a well-formed proposal from a single practitioner

source, not as a quoted industry standard** (§5 provenance; the two GDC talks are pipeline

references and were not verified to contain this protocol).

• **Adopt Loomis, Wikipedia's canon summary, and the "hand = face / foot = head / legs = 3.5–4 heads"

rules as SANITY RAILS ONLY**, with their measured error at A14 recorded on the face of the doc

(+7%, +12%, and 3.48 HU respectively). They are useful as coarse floors and are wrong as targets.

• Do not adopt any number from a source that does not declare its convention, and do not adopt

the ellipse shortcut for chest depth under any circumstance.

---

9. GAP NOTES — where our named stack already deviates from the practice above

Written for the audit phase. Each item names the file, the deviation, and the arithmetic or the

citation that establishes it. Nothing here is a style preference.

GAP-1 · The acceptance band is the disagreement between two of our own specs, and it is enormous

harness/asset_factory/thousand_point_eval.py (docstring, lines ~55–62) defines the PROPORTION band

as "the disagreement between the two grounded sources this programme already carries for that site,

BODY_AGE_NN.json and BODY_AGE_NN_RESEARCH.json, floored at 2 percent," and the ANATOMICAL band as

"the interval spanned by those same two sources, widened by nothing."

Those two files disagree, at A14, by:

A mesh may be 20% wrong at the hips and 13% wrong at the shoulders and be scored inside band. That

is a sufficient mechanism, on its own, for "shipped past every gate." Worse, the relationship is

backwards: the sites where the research *corrected* the landed spec get the widest bands, while

the sites AGE_PROPORTION_RESEARCH.md §6 explicitly declared **unreached and carried through

unchanged** — underbust, knee, elbow, hand_circumference, plus thigh_length, shank_length,

elbow_height, wrist_height, foot_breadth, hand_breadth — are identical in both files and so

collapse to the 2% floor. The least-grounded numbers get the tightest band and the best-grounded

get the loosest. Fix is R-7.

GAP-2 · The landed A14 spec still carries the defect its own research diagnosed

build/3d/characters/body_specs/BODY_AGE_14.json :: segment_lengths_cm.biiliac_breadth = 31.844 is

0.19100 of stature. AGE_PROPORTION_RESEARCH.md §1 identifies exactly that value — "0.1910 at

age 7 and 0.1910 at age 18," a flat fraction traced to the engine's own

FHumanityChildProportionSpec defaults — as the root-cause artefact, against a truth of

0.1546→0.1562. The research file has the fix (25.992 = 0.15590); the landed file does not. Both are

live, and consumers disagree about which to read:

• build/3d/best_a14/bounded_fit.py:76 → BODY_AGE_14_RESEARCH.json
• build/3d/best_a14/silhouette_overlay.py:75 → defaults to BODY_AGE_14.json
• harness/asset_factory/thousand_point_eval.py:2741-2742 → reads both and bands between them

One of these is the target and the other is history. Until that is ruled, every downstream number is

ambiguous by up to 20%.

GAP-3 · No limb-girth lens exists, and limb girth is where gaunt lives

AGE_PROPORTION_RESEARCH.md §4 tabulates chest, hip and neck girth and **no arm, thigh or calf

girth series**; §6 names knee, elbow, underbust and hand circumference as unreached. The specs do

carry bicep_relaxed, thigh_proximal, calf_max, forearm_max, wrist_min, ankle_min — but

four of those (knee, elbow, underbust, hand_circumference) are ungrounded carry-throughs

(GAP-1), and nothing in the lens battery scores a limb girth against a population reference. A figure

can be exactly right in every length and breadth and still be a stick. [KRK] gives the population

distribution for the fat compartment (§3.5) and is the missing reference.

GAP-4 · The leanness knobs have no floor and no re-derivation

BODY_AGE_14.json :: training_clock.adiposity = -0.45, surface_definition_index = 0.3,

subcutaneous_fat_scalar = 0.8995, slenderness_vs_isometric = 0.96911. Nothing clamps these against

a measured distribution, and nothing re-derives the implied body fat from the mesh that comes back.

The knobs are trusted as inputs and never audited as outputs. [KRK] body fat 20.0 ± 11.6% is the

clamp. Fix is R-12.

GAP-5 · The A18 body solve reads the un-corrected spec, and reads it in the wrong convention

build/3d/characters/body_specs/BODY_ROUTE_A_V8_A18.json records

spec_head_cm_over_shoulder_cm 0.4721 against shoulder_px_measured 370 at

px_per_cm_from_shoulders 7.603. Solved out, that is an assumed shoulder of 48.665 cm and an

assumed head height of 22.975 cm — which are, to three decimals,

BODY_AGE_18.json :: across_shoulder_soft_tissue 48.663 and head_height_cm 22.974. The solve is

consuming the landed spec, not the research one, which is GAP-2 with a second consumer:

So the solve runs +5.8% on head units and +5.2% on shoulder width against the corrected ladder,

and it lands beside the seven-age lineup's own diagnosed D-3 error at the same rung (+12.6%

bideltoid at 18). Underneath that is Trap 1: [SNY] puts the age-18 menton→scalp head at

22.13 cm, so the landed 22.974 is a near-scalp figure, while the solve measures crown of mesh,

which is hair — and the hair-convention target is 24.073 cm at 7.5 HU. Half a head unit of

convention error, in the direction that makes the head too small and therefore the shoulders read

broad (R-3). The file should state, per number, which spec and which convention it is in.

GAP-6 · anatomy_data.py and check_anatomy_conformance.py have no HUMANOID proportion content

harness/asset_factory/anatomy_data.py (111 KB) is comparative-animal anatomy — the eight bucketed

creature archetypes, joint ROM, gait. check_anatomy_conformance.py scores bone roster, segment

count, required bones, ROM and centreline: presence, count and range — arithmetic, and it says so.

Neither carries a single human proportion, girth, landmark or fat figure. There is no file in the

harness today that would fail a gaunt human, which is consistent with one shipping. anatomy_bar.py

is that file and should sit beside them with the same citation discipline anatomy_data.py already

enforces ("a degree figure with no citation key is a defect in this file") — extended to

"a proportion, girth or millimetre figure with no citation key is a defect in this file."

GAP-7 · The silhouette gate is one framing, where the practice is three distances and five poses

build/3d/best_a14/silhouette_overlay.py compares against a spec at a single framing. The production

protocol in §5 is a black-silhouette recognition test at hero, gameplay and recognition-threshold

distances across five canonical poses, with blockout failure returning the asset to blockout. A

single mid-distance overlay cannot detect the class of error that only appears at 50 m or only in an

attack pose. Fix is R-9.

GAP-8 · Landmark instruments are measuring different rulers, and it is already known

AGE_PROPORTION_RESEARCH.md §5 records, under NOT SCORED, that the diagnostic instrument's thigh

(crotch → the leg's own girth minimum) and shank (that minimum → the ankle's minimum) are not the

research's trochanterion–tibiale and tibiale–sphyrion, and that its head breadth includes the cage

head's hair — the rows differ by 23–66% and none of it is a defect measurement. That reconciliation is

owed work and is a prerequisite for anatomy_bar.py, which will otherwise inherit the same two-ruler

problem at three sites on day one. R-1 is the structural fix; the reconciliation itself is the task.

GAP-9 · The independent craniofacial cross-check is still unread

AGE_PROPORTION_RESEARCH.md §6 names Farkas, *Anthropometry of the Head and Face* (1994) Appendix A

as located, confirmed to hold the right table, and lending-restricted. It remains the single most

valuable outstanding cross-check on the head curve — from an independent sample and an independent

method — and the head curve is the reference the whole shoulder read depends on (§3.2, R-3). Under

the standing books-never-gate ruling it does not block; it lands as a depth pass. Recorded here so it

is not lost between two documents. The CDC Series 11 No. 249 biacromial table is in the same state:

the reference standard, HTTP 403 to every automated fetch tried from here, on this attempt as on the

last.

Generated by harness/site/structure_site.py — the URL path is the repo path. review root